Ranking and freshness

Versioned methodology

The global rank is a deterministic output of an active methodology version and completed cohort run. A result records eligibility, component values, normalized contribution, weight, coverage, freshness, confidence, and source watermark. Missing evidence lowers coverage/confidence; it is not silently treated as zero performance.

Rank activation is atomic. Readers see either the previous complete generation or the new complete generation, never a partial cohort. Daily snapshots preserve score, rank, one-day movement, and seven-day movement for history and repair.

Reading a score

• scorePpm is the methodology output in integer ppm.
• rankConfidencePpm describes evidentiary confidence, not probability of success.
• projectedAt identifies the directory projection time.
• methodology code/version and ranking completion identify the cohort basis.
• a null rank means ineligible or insufficient data, not last place.

Sponsored promotion can influence labeled discovery surfaces only according to its published policy. It never changes the organic ranking result.

Freshness and repair

Ranking consumes projection watermarks rather than making provider calls on the request path. Shadow runs can be compared before activation. Consistency checks, full rebuilds, entity repair, and daily publication are durable idempotent jobs. A repair appends audit evidence and activates only validated output.

Use the methodology page for human-readable component policy and the rank/history API for exact snapshots. Do not compare scores produced by different methodology versions as if they share one scale without an explicit migration analysis.